Using reinforcement learning for engine control

The experiments described are directed towards using reinforcement learning to solve control problems for a combustion engine. The control task presented is to follow an arbitrary sequence of target values for the number of revolutions under the additional condition of keeping the air-to-fuel-ratio close to the optimum by manipulating the system inputs throttle valve angle and fuel injection duration. For this challenging problem of controlling a nonlinear multiple-input-multiple-output system an autonomously learning multi-controller architecture is developed. We also present a comparison to conventional approaches using PI-controllers developed according to the frequently used Ziegler-Nichols parameter adaptation rules.